Your next best customer is a machine

Try this before you read any further. Open an assistant, give it a normal weekly shopping list, and ask which supermarket is cheapest. It will do a competent job of the visible part: parsing the list, resolving the products, coming back with three totals and a recommendation. What it won’t tell you is that the totals are wrong, because it quoted the price you would pay walking in without a loyalty card, which almost nobody does. The Competition and Markets Authority puts scheme membership at 97% of shoppers, holding three schemes each on average.

The assistant is not malfunctioning. It read what the retailers published, and what the retailers published was the wrong number.

Notice where it went wrong. Not at discovery: the agent found every retailer and resolved every product without difficulty. The failure came afterwards, at the point where the agent had to work out what buying from you would actually cost and how it would be done.

That distinction matters because it is a different problem with different owners and a different fix, and it is the one this piece is about.

The short answer is: the least useful version of you.

FIGURE 1 — WHAT RESOLVES WITHOUT A SESSION, AND WHAT DOES NOT

Above the line sits the public surface, which is thin. Product name, standard price, an image, some identifiers, a category. That is roughly everything a machine can reach without credentials. Below it sits the member price, the weekly unlocked offers, basket-level promotions, points balances, store-level stock at a named branch, delivery slot availability, and the price match adjustments that only resolve at the till. Every item in that second list represents years of investment, and most of it is what a retailer would name if you asked what makes them competitive.

Once you see the pattern, it’s pretty consistent. Everything built to make an offer personal was built as a digital experience to be drawn on a screen for a signed-in human, so none of it survives contact with an agent.

Even the public part is answering the wrong question. Feeds were designed to match a query to a SKU for an ad auction, so they carry what an auction needs. An agent is doing something else. It is reasoning about a decision, which means it wants to know whether the coat packs down small, whether the part fits a 2019 model, how the sizing runs, and whether the thing will actually arrive before Saturday. Almost none of that is published in a form a machine can use, which is why agents fall back on the attributes that are. If the only axes you publish are price and availability, price and availability are what you will be judged on.

Let’s start with price, usually it’s the easiest to quantify.

Over the last four years the major grocers rebuilt pricing around membership, to the point where the shelf carries two numbers and only one of them is real for most customers. The CMA looked at roughly 50,000 loyalty-priced products across Tesco, Sainsbury’s, Waitrose, Co-op and Morrisons and found average member savings between 17% and 25%.

FIGURE 2 — THE GAP BETWEEN THE PUBLISHED PRICE AND THE PRICE PAID.

Compare that against your own net margin for a second. Nobody in this sector wins share on twenty points. They win it on fractions of one. And unlike a lost search ranking, you get no signal that it happened, because the comparison completed and the customer got an answer.

The surprising part is that the fix already exists. Google added MemberProgram and MemberProgramTier to its structured data support in June 2025, and an offer can carry a UnitPriceSpecification with validForMemberTier attached, which is a boring way of saying a page can declare a non-member price and a member price side by side, as data, with the tier they belong to. That is live in UK search results today.

So the mechanism shipped over a year ago and almost nobody has claimed it.

When we raise this with retail teams the reaction is usually not disagreement so much as a slightly awkward pause, because loyalty pricing sits with commercial and structured data sits with digital, and the two have never had a reason to talk.

Publishing better data gets you a correct comparison. It does not get you a completed basket, and the second problem is harder because it is not really a data problem.

For an agent to buy on a customer’s behalf, it needs to prove who that customer is without being handed the whole account, which means scoped and revocable credentials rather than a password and a session cookie. It needs the offer to be an object it can hold, with an identity and an expiry, instead of a condition a page happens to be in when it loads. It needs promotions to resolve at quote time rather than at checkout, a change to the AI workflows and automation behind the basket, not to the pricing page or the number it gives the customer is wrong even when it is fully authenticated.

Each of those is a decision about how much authority you delegate, and the one I would think hardest about is substitution.

In grocery, out-of-stock handling is currently a retailer algorithm running on retailer margin logic. When an item is unavailable, you decide what goes in its place, and a meaningful share of the time what goes in its place is your own brand. Own-label is the margin engine, and substitution is one of the quiet mechanisms that feeds it. The moment an agent assembles the basket, that decision moves, unless you have deliberately designed substitution as something the agent negotiates with you rather than resolves on its own with a generic preference for cheapest-equivalent.

The tension here is that agents optimise for the shopper, so clear delivery windows, published shipping costs and a machine-readable returns policy all make you more selectable, and the last one makes you more expensive to serve. Vagueness used to be commercially useful, because a shopper already half committed on your page will tolerate “delivery from £3.95” and find out the real number at checkout. An agent comparing four retailers reads that as missing data and moves on. Legibility and margin genuinely pull against each other here, and the retailers who decide where they want to sit will do better than the ones who arrive at a position by accident.

Retail media spend in this market is expected to top £4.8bn this year against roughly £4bn in 2025, and Nectar360 on its own is targeting at least £100m of incremental profit by March 2027. That business runs on a logged-in human looking at a rendered digital experience. Sponsored placements and shelf position and onsite impressions all assume a browsing session, and the browsing session is precisely what an agent takes out of the journey.

Meanwhile the investment is flowing the other way. Sainsbury’s now asks customers to unlock Your Nectar Prices in the app before they shop, refreshed every Friday, behind a decisioning engine working towards 500 million personalised offers a week. Tesco is heading the same way through its Adobe partnership across 24 million Clubcard households.

Don’t think of this as a mistake, because it really isn’t. Personalisation is the most defensible asset the big grocers have. But the money is compounding inside the layer agents cannot read, at the exact moment agent adoption is accelerating. Adyen’s 2026 Retail Report has assistant use among British shoppers at 28%, up from 12% a year earlier, with 84% of retailers open to letting AI complete purchases for a customer. And the thing shoppers most trust an agent to do is find the best price, which ACI’s YouGov work put at the top at 50%, comfortably ahead of trusting an agent to act in your financial interest or sort out a problem.

Loyalty pricing and retail media look like separate problems on an org chart. They share a dependency.

Treat agentic commerce readiness as a gradient rather than a yes-or-no, because transactable is not a state anybody is simply in or out of.

FIGURE 3 — FOUR STATES OF AN OFFER, FROM RENDERED TO TRANSACTABLE.

• State 0, rendered. The price exists only inside a page, drawn for a signed-in human.
• State 1, declared. The standard price is published as structured data and in the feed. Most retailers sit here.
• State 2, conditional. The member price is declared as data, tied to a named programme tier.
• State 3, transactable. An agent can hold that offer, build a basket and complete against it.

State 2 is a publishing job. A retailer that wanted to could plausibly have it done inside a quarter, and whoever does it first spends the following few years being quoted correctly in comparisons where their competitors are not. It is the cheapest competitive advantage currently available in this sector, which is a strange sentence to be able to write.

State 3 is where the authority questions land, and nobody has settled them. If you hand an agent an authenticated session, who owns the customer relationship afterwards? Does the agent pass the member discount to the shopper or bank it inside its own comparison? What happens to the value exchange that justified collecting the data in the first place? I don’t think anyone should pretend to have clean answers yet, but the retailers building against those questions will end up with better ones than the retailers waiting to be told.

The reassuring part is that nobody is behind on the hard bit. Multi-item basket checkout for agents does not exist in grocery in any market, which Ocado has said publicly, noting that single-product off-platform checkout works while multi-product does not. The easy part, though, has been sitting there for over a year.